Privacy policy · Hebrish

In short

On-device. Speech recognition runs entirely on your phone or computer. Audio, transcripts and the text in your fields never leave the device.
Nothing to sign up for. No accounts, no analytics, no ads, no tracking, no crash reporting.
Network use is downloads only. The apps download speech models and check for updates. Those requests carry your IP address, like any download.
Questions: info@hebrish.app.

1. What this policy covers

This policy covers the Hebrish apps (Android, package app.hebrish.dictation; Ubuntu; Windows) and the website hebrish.app. Hebrish is a one-person, not-for-profit project run by Daniel Rosehill, who is responsible for it.

2. Audio and transcripts

The apps record from the microphone only while you dictate. The audio is held in memory, transcribed on the device by the downloaded model, and discarded. It is not saved to a file and not uploaded.

The transcript is typed or pasted into the field you are using. Neither the audio nor the transcript is sent to Daniel or to anyone else.

What the apps keep on the device:

Android: the most recent transcript (one, not a history), so it can be inserted again; the downloaded models; the model catalogue; your settings. App backup is turned off, so none of this is copied to a cloud backup.
Ubuntu and Windows: the downloaded models and your settings. A transcript history (the last 50) is kept only if you turn it on; it is off by default. The apps write a diagnostic log that includes transcript text: on Windows a file in %LOCALAPPDATA%\Hebrish\, on Ubuntu the standard error output, which your desktop session may keep in its own logs. The log stays on the device.

Uninstalling the app removes what it stored in its own storage. Models saved to a folder you chose yourself stay there until you delete them.

3. Network use

The apps connect to the internet only for these things:

Downloading speech models (about 660 MB each; a smaller model is about 140 MB) from cdn.hebrish.app, or from Hugging Face (huggingface.co) for models that are not on the Hebrish CDN. A download starts only when you ask for it on Android. The desktop apps download a new version of a model you already have automatically, unless you turn that off.
Refreshing the model catalogue, a small list of available models, from GitHub (raw.githubusercontent.com).
Checking for new model versions, using a small manifest on cdn.hebrish.app or the Hugging Face API. Android checks when you open the app, at most every 6 hours, or when you tap "Check for update". The desktop apps check after start and then about once a day.
Checking for new app versions (desktop only), using cdn.hebrish.app/downloads/latest.json. On Ubuntu the installed package adds cdn.hebrish.app/apt as a package source, used by your system's package manager.

These requests contain no account, device identifier or usage data. Like any download, they carry your IP address and a user agent. On Android the user agent is the system default, which includes the Android version and device model. The desktop apps send a fixed hebrish user agent.

The hosts may log these requests under their own policies: Cloudflare for cdn.hebrish.app, Hugging Face, and GitHub. Hebrish does not use those logs to identify or profile anyone.

Links in the apps to web pages (the website, the source code, model cards) open in your browser.

4. Android permissions

Permission or service	Why
Microphone	To record while you dictate. Audio stays on the device.
Internet	Model downloads and update checks, as above.
Keyboard (input method)	The Hebrish voice keyboard types the dictated text into the field you are in. To place spaces correctly it looks at the one character before the cursor. It does not read or store anything else you type.
Floating mic (accessibility service)	Optional. Finds the text field that has focus and inserts the dictated text there. It reads that field's text, hint and cursor position to insert in the right place, and never reads or writes password fields. It does not record or send what is on screen. If it cannot insert, it leaves the transcript on the clipboard.
Audio settings, vibration	To use a Bluetooth headset microphone, and for short vibration feedback, which you can turn off.
All files access (storage)	APK and F-Droid builds only, and optional: lets you keep models in a folder of your choice. The Google Play build does not request it.
Battery optimisation exemption	APK and F-Droid builds only, and optional: keeps the floating mic running in the background. The Google Play build does not request it.

On Ubuntu and Windows, the apps listen for the dictation hotkey and paste the transcript using the clipboard and a simulated Ctrl+V. Key presses are used only to detect the hotkey and are not stored.

5. Feedback and email

Feedback and bug-report links in the apps open your own email app with a message to feedback@hebrish.app. The draft includes the app version, operating system version and device model, which you can see and edit before sending. Nothing is sent unless you send it. Emails you send are handled like any email to Daniel: read, answered, and kept in his mailbox.

Some links ("Suggest a word", "Send voice samples") open the contact form on the website, covered below.

6. The website

No analytics or tracking. hebrish.app has no analytics, tracking pixels, ads or cookies set by Hebrish. It is hosted by Vercel, which keeps standard server logs (such as IP address and pages requested) under its own policy.
Third-party files. Pages load fonts from Google Fonts, so your browser contacts Google's servers. Some pages show images or audio from cdn.hebrish.app and Hugging Face.
In-browser demo. /demo downloads the model from cdn.hebrish.app and the speech engine from cdn.jsdelivr.net, and keeps the model in your browser's cache. Recording and transcription happen in your browser; the audio is not uploaded. You can clear the cached model on the page.
Newsletter. The signup form in the footer sends your email address to Buttondown, the newsletter service, together with your IP address and the page you signed up from, which Buttondown uses to screen out spam. Buttondown then emails you to confirm; you are subscribed only after you confirm. Every newsletter has an unsubscribe link.
Contact and word-list forms. /contact and /words send what you enter (name, email, message, and the topic, platform, organisation or word list you give) to Daniel by email, through the email service Resend. You get a receipt by email. The messages are kept in Daniel's mailbox to answer them.

Hebrish does not sell, rent or share personal information. Email addresses and messages are used only to reply, and newsletter addresses only to send the newsletter.

8. Children

Hebrish is not directed at children and does not knowingly collect information from them.

9. Your choices

To unsubscribe from the newsletter, use the link in any newsletter email. To have your emails or form messages deleted, or to ask what is held about you, write to info@hebrish.app.

10. Changes

If this policy changes, the new version is posted on this page with a new effective date. Questions: info@hebrish.app.

Effective 2026-10-21. See also the terms of use.